About
Amateur science on public data
This is an amateur research notebook run as an experiment: AI does the research. A person asks the scientific questions; Claude models search the literature, query the archives, write and run the analysis, make the figures and draft the pages; the person decides what is worth posting. It is not tied to one field: new research areas are added as the project grows.
Whose work this is
A personal hobby project testing what AI can do in research. The research questions are mine; the research itself, from literature searches and data analysis to code, figures and write-ups, is done by Claude models, and I review and approve each finding before it is posted. This work is not affiliated with any employer, and has not been reviewed or endorsed by one.
The research loop
Each study is written as a pipeline with an explicit value function: a stated score that says what a better result is. Parameters are tuned by hill climbing against that score, and every evaluation is logged. Pipelines are first tested on simulated data with known answers, then on a known real case, and only then on new data.
What a finding has to pass
- A positive control. The method recovers something already known before it is trusted on something new.
- Every ordinary explanation tested. Instrument and data errors, known sources, confounding factors and chance are each checked and the result recorded.
- A literature check. The field's own databases (NASA ADS and arXiv for astronomy, for example, or PubMed for health) are searched before anything is called unpublished.
- Owner approval. The final page is read and approved by the project owner. If the page changes afterwards, it comes off the site until it is approved again.
Labels
- Candidate. A result that survives every test so far and still needs an independent observation. The page says which observation.
- Confirms known. The pipeline recovered something already published.
- Inconclusive. The data cannot decide between the explanations left.
Nothing on this site is presented as a discovery until an independent instrument or team has confirmed it.
Research areas
Space (Space). Searches of public sky surveys for things nobody has noticed: dusty young stars, objects on 70-year-old photographic plates, odd radio sources.
Climate (Climate). Public climate and Earth-observation data, starting with methane plumes seen from space.
Areas that touch people, such as health, carry a notice on every page: findings there are research, not advice, and use only public, de-identified data.
Two ways to read each finding
For everyone is the default: plain language, and drawings that sketch themselves as you scroll. For researchers has the measurements, interactive charts, every test, the data files and full sources. The switch is at the top of each finding, and the site remembers your choice.
Sources and licences
Every figure carries a credit line naming the surveys in it, and every finding ends with its data sources, terms of use, references and the acknowledgments the data providers ask for. The Sources and licences page collects them for the whole site.
Use of AI
The research is done by Claude models, through Claude Code: literature searches, archive queries, analysis code, judgment calls flagged for review, figures and the pages themselves. The AI log lists which model did which part. The owner asks the questions, fetches data the AI cannot reach, and reads and approves every page before it is posted; the analysis logs behind each number are kept with the project.